Claude Code + Clay Lead Generation: From Plain-English Brief to Reviewed Outreach

The direct answer

Nate Herk's Claude Code and Clay demonstration is a strong example of agent orchestration applied to sales operations. Claude Code receives a plain-English outcome, Clay supplies company and contact data through its MCP connection, a waterfall searches and validates work emails, and the agent produces a structured lead file with research and draft copy.

The valuable idea is not that one prompt replaces prospecting. It is that the workflow can be expressed as a reviewable contract: define the market, gather evidence, enrich conservatively, reject weak records, draft from real business context, and stop before sending so a person can review the list and message.

Source Note: a useful demo, not a conversion promise

Nate reports that the detailed HVAC run found 50 leads, launched six research agents, took about an hour, and consumed 172 Clay credits, which he described as roughly $12. A simpler 50-lead request reportedly took about five minutes. Those figures describe his July 2026 demo configuration, not a guaranteed price or speed.

The transcript also estimates that a waterfall might raise email coverage toward 80% to 90%. I have not reused that as a general fact. Clay's official documentation supports the mechanism - sequential providers, configurable validation, and stopping on a valid result - but coverage varies with the market, geography, source quality, provider order, and definition of valid.

Clay's current pricing separates Data Credits from Actions. Searches, enrichment providers, AI research, phone data, sequencing, and retries can consume different resources. Always run a small sample and inspect the actual usage report before projecting campaign economics.

Link Map

The architecture: Claude orchestrates, Clay executes data work

Nate separates the problem into two layers. Clay addresses fragmented data sourcing and enrichment. Claude Code addresses the tool problem by turning the workflow into a natural-language request and coordinating the steps.

1. Brief: define the company profile, target account, decision-maker role, geography, exclusions, offer, proof, and desired next step.
2. Source: ask Clay to find candidate accounts and contacts from approved data sources.
3. Enrich: append company facts, role, website, work email, phone where justified, and source URLs.
4. Validate: check email status, deduplicate records, screen suppression lists, and reject weak matches.
5. Research: gather a source-backed business signal that makes the offer relevant.
6. Draft: generate a subject and body from the business context, not from generic AI filler.
7. Review: deliver a CSV and run summary. Do not send automatically.

Clay's official MCP guide describes the connector as exposing more than 150 data providers, AI research agents, and prebuilt workflows inside supported assistants. It also makes an important organizational distinction: operations teams design and govern workflows; representatives consume those vetted capabilities in chat.

Why the waterfall is useful

No contact-data provider has perfect coverage and accuracy. A Clay waterfall queries providers in an ordered sequence and stops once a result satisfies the configured validation rule. For work email, Clay can optionally infer a likely address first, validate it, and only call paid providers if necessary.

The operator still decides what counts as acceptable. Clay documents conservative, balanced, aggressive, and advanced validation strategies. For cold outreach, conservative validation is the safer starting point because catch-all and uncertain results can increase bounce and reputation risk.

Personalization begins before the lead search

Nate's demo project includes a business profile, website copy, offer, FAQ, proof, and case studies. This context is not optional decoration. Without it, the agent knows the prospect but not why the sender deserves a reply.

A useful project context folder should contain:

• A one-page ideal customer profile with explicit exclusions.
• The offer, outcome, price range, delivery constraints, and who should not buy.
• Approved proof with exact claim boundaries and source links.
• Case studies that state the client, problem, intervention, and measured result.
• Common objections and honest answers.
• Brand voice examples and phrases to avoid.
• The permitted call to action and follow-up policy.
• Legal, regional, suppression, and sensitive-data restrictions.

Personalization should explain relevance, not display surveillance. A recent company announcement, public hiring signal, service page, or clearly sourced review pattern may support a useful hypothesis. Private details, sensitive attributes, invented pain points, or unnervingly specific personal observations should not appear in outreach.

A safer goal prompt for Claude Code and Clay

Nate's prompt requires 50 fully enriched leads with no blank columns. That produces a clear completion target, but it can also reward the agent for filling gaps rather than admitting that evidence is missing. This version preserves the outcome while making uncertainty and human approval part of the contract.

I need a review-ready prospect list for [offer].

Target:
- Industry: [industry]
- Geography: [region]
- Company criteria: [size, business model, signals]
- Decision-maker roles: [roles]
- Exclude: [competitors, customers, suppressed domains, risky categories]

Read the project context before starting:
- business profile
- offer and delivery limits
- approved proof and case studies
- brand voice
- compliance and suppression rules

Use Clay to find up to 50 qualified prospects. Quality is more important
than reaching 50. Return fewer records if the evidence is weak.

For every accepted lead, include:
- company and website
- contact name and role
- validated work email and validation status
- source URL for every material claim
- one recent, business-relevant signal
- a short explanation of offer fit
- a personalized subject and plain-text draft
- review flags and uncertainty

Rules:
- Do not invent missing fields, pain points, achievements, or proof.
- Do not use sensitive personal data.
- Deduplicate by company, person, domain, and email.
- Screen the suppression list before acceptance.
- Use conservative email validation.
- Estimate credits before the full run and test 10 rows first.
- Do not create a campaign or send any message.

Deliver:
1. review-ready CSV
2. rejected-record CSV with reasons
3. source and validation report
4. credit and action usage summary
5. five sample emails for human review

Stop and ask before spending above [budget] or changing the target.

The reported $12 demo and real unit economics

Nate reports 172 Clay credits and describes the run as about $12. Treat that as a field note, not a price list. Clay's current model uses Data Credits and Actions, and the actual cost changes with contact providers, phone enrichment, AI web research, validation, sequencing, retries, and plan terms.

Track three costs separately:

• Cost per sourced candidate: every account or contact returned before quality review.
• Cost per accepted lead: only records that pass fit, source, validation, duplicate, and suppression checks.
• Cost per positive outcome: replies, qualified conversations, proposals, and closed work, not opens.

Nate uses a hypothetical 1% conversion to illustrate potential return. Do not put that assumption into a forecast until your own campaign produces enough evidence. A cheap list can still be expensive if the offer is weak, replies are negative, or the sending domain is damaged.

Verified data is not verified relevance

The HVAC CSV in the video includes business name, decision maker, title, email status, phone, website, city, reviews, a proposed pain point, a recent signal, a personalization hook, source, subject, body, and review notes. That is a useful review schema.

Before accepting a row, check:

• Does the person still work at the company and plausibly own this problem?
• Does the source actually support the claimed signal?
• Is the email valid under the chosen validation strategy?
• Is the recipient an individual, sole trader, partnership, or corporate contact under the relevant rules?
• Has the person or company opted out before?
• Does the sender's proof support the promise in the draft?
• Would the email still feel relevant if the personalization line were removed?

In Nate's example, the draft references a negative review about missed callbacks. That can be relevant to the offer, but it needs care. Confirm the review is current and accurately represented, avoid shaming the business, and frame the observation as a hypothesis rather than a diagnosis.

Launching in Clay should be a separate decision

The video imports the completed CSV into a Clay table, maps subject and body variables, previews the messages, and demonstrates sender-account and campaign setup. Clay's current sequencer documentation confirms campaign schedules, message previews, per-lead edits, warmup, send limits, analytics, replies, blocklists, and campaign event tables.

The generation workflow and the sending workflow should remain separate. A successful CSV export must not automatically become permission to contact everyone in it.

1. Review at least the first 10 records and every high-risk flag.
2. Send test emails to internal addresses and inspect missing variables.
3. Confirm sender identity, physical/contact details where required, and opt-out handling.
4. Apply the global suppression list before launch.
5. Use a small first batch and conservative daily limits.
6. Watch bounces, complaints, negative replies, and qualified replies.
7. Pause when quality deteriorates; do not compensate by adding volume.

Compliance and privacy boundaries

This section is operational guidance, not legal advice. Rules depend on the sender, recipient type, jurisdiction, data source, relationship, message, and communication channel.

The European Commission says an organization acquiring third-party contact data must be able to establish that the data was obtained lawfully and may be used for advertising. It must keep the data current, respect objections, comply with ePrivacy rules, and provide required transparency no later than the first communication.

The ICO's operational guidance distinguishes individuals, sole traders, some partnerships, and corporate bodies, and emphasizes sender identity, a valid opt-out method, and a do-not-contact list. The FTC's CAN-SPAM framework applies to US commercial email. These regimes are not interchangeable, so obtain jurisdiction-specific advice before scaling across regions.

The JQ AI SYSTEMS operator workflow

A seven-day validation plan

1. Day 1 - Define: choose one offer, one buyer, one region, and explicit exclusions.
2. Day 2 - Context: prepare approved proof, case studies, voice, compliance rules, and suppression data.
3. Day 3 - Connect: configure Clay and Claude Code using the current official setup path; set a small credit limit.
4. Day 4 - Test: source and enrich 10 records using conservative validation.
5. Day 5 - Review: reject weak rows, verify every signal, and rewrite five messages manually.
6. Day 6 - Pilot: launch only a legally reviewed, tightly bounded batch with working opt-out and suppression.
7. Day 7 - Measure: record accepted-lead cost, bounce rate, objections, positive replies, and learning for the next batch.

The goal is not 500 leads overnight. It is a small system that can explain where every record came from, why the person fits, what the message claims, what it cost, and who approved the contact.

Sources

• Nate Herk: Claude Code + Clay Makes Lead Generation Actually Fun
• Nate Herk on X
• Clay official website
• Clay: What Is Clay MCP?
• Clay Docs: Work Email waterfall
• Clay Docs: Email sequencer
• Clay current pricing and plan limits
• Anthropic: Claude Code plugin marketplaces
• FTC: CAN-SPAM Rule
• European Commission: third-party data used for marketing
• ICO: electronic mail marketing guidance